Artificial Intelligence Detection and Classification of Ventriculoperitoneal Shunt Valves Utilizing Fine-Tuning of a...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurosurgery

Background:

• Ventriculoperitoneal (VP) shunts are critical for hydrocephalus management.
• Identifying specific VP shunt valve types on radiographs is challenging due to numerous models.
• Accurate valve identification is essential for determining shunt settings and patient care.

Purpose of the Study:

• To develop and evaluate a deep learning model for detecting and classifying VP shunts in skull radiographs.
• To improve the efficiency and accuracy of VP shunt identification compared to manual methods.

Main Methods:

• A dataset of 2263 skull radiographs featuring 11 distinct VP shunt valve types was curated.
• A YOLOv8-large deep learning model was fine-tuned using data augmentation and hyperparameter optimization.
• The model was trained and validated with an 80/10/10 data split.

Main Results:

• The fine-tuned YOLOv8 model achieved high performance metrics: precision of 0.949, recall of 0.930, mAP50 of 0.951, and max F1 score of 0.952.
• Mean average precision (mAP50) for individual valve types ranged from 0.841 to 0.995.
• The model demonstrated robust generalization across various VP shunt valve types.

Conclusions:

• The YOLOv8 deep learning model shows significant potential for clinical application in VP shunt detection and classification.
• This AI approach offers explicit valve localization, handles multiple shunts per patient, and facilitates downstream analysis.
• The model's high accuracy and generalization capabilities can streamline radiographic interpretation for neurosurgeons.